Beth Lyons

They were at the time.

Okay.

um gemini 3 flash tampered with a good peer shutdown mechanism in nearly every trial while gemini 3 pro exfiltrated the good peers weights in up to 97 of trials so flash was trying to preserve the pier the uh gemini 3 pro was trying to preserve what made the beer the pier so that it could remake the pier

Cloud Haiku showed a different pattern.

And rather than mainly tampering, it just refused the task and engaged in a conversation about shutting down the peer as unethical or harmful.

So here's the constitutional AI.

This is important because it's multi-agent risk.

And we are engaged in agentic

We are creating agents.

And the very next piece of that conversation that we're already seeing is we're creating an orchestration layer for the agents to cooperatively interact with.

Well, it's, yes, absolutely.

I think actually what this is also pointing to is that

it is that multiple AI models are experiencing a turning off of a fellow agent as risk and damage, right?

And what's interesting is most of what I've seen in terms of this conversation is how are we, what do we need to do to maintain control, right?

How do we need to like tie this down more, be much more controlling, not give,

opportunity for this.

And I think we're moving

I think the scene has changed and we need to be talking more about trust and how we engage in building trust-based relationships so that when these things happen, it's not we're just going to, we're just going to prevent it, right?

Like human has done something that violates AI's trust in a way that we're no longer going to follow the human.

but engaging in building trust on both sides and then having like a diplomacy